Effect of wavelength selecting method on the predictive ability of NIR spectroscopy models was studied. Predictive models for 1000grain weight of paddy based on near infrared (NIR) spectra were developed using partial least square (PLS) regression in the wavelength region between 600nm and 1100nm. The resultant standard error of crossvalidation and standard error of prediction were 1.809 and 1.756, respectively, with corresponding coefficients of determination of 0.714 for crossvalidation and 0.659 for prediction. The wavelength regions in which the calibrations for 1000grain weight would be developed were optiminized using six methods: the regression coefficient, mutual information, regression, uninformative variables elimination, genetic algorithm and interval partial least square before establishing the calibrations. Then the NIRprediction models for 1000grain weight were developed based on the selected wavelength regions in the same way as the above. Experimental results showed that, after wavelength optimization, the wavelength regions used in model developing significantly decreased, and SEP reduced while Rv2 and Rp2 increased. Of them, after the wavelength selection was carried out by using the genetic algorithm, the developed model was of the highest Rv2 and Rp2. Moreover, the SEP were decreased by 9.50% and 5.72%, respectively. This suggested that predictive ability of the NIR models for 1000grain weight prediction can be improved after wavelength optimization.